Listen: https://serve.buildtall.systems/16045e92ea7e4adf71c15f20d7a66dd5772dacd45d95ef471b25c2a273c002e3.mp3
Source: https://seroter.com/2026/11/17/daily-reading-list-november-17-2026-911/
In today's Seroter's Daily Reading, the post is Daily Reading List – November 17, 2026 (#911). Richard Seroter opens by noting it's another long list, and his reading queue remains stubbornly long, so we'll get through it.
The first item is The levels of agentic engineering (from someone who’s led the climb twice). Seroter says forward progress isn't just about slapping on another set of tools; this author looks at a maturation of workflow. He's not sure about the last few stages, because identifying the right work should come sooner, but he finds it a useful framing. The author, Zach Wills, has led an AI transformation twice, and instead of a ladder of tools, he asks one question: what are you still doing yourself? The seven levels move from handing the ticket to an agent, to making the codebase explain itself, letting the agent prove its own work, getting the work off your laptop, closing the loop in production, making the factory multiplayer, and finally deciding what's next. Most teams stall at the cloud step, and the tools keep changing, but that central question holds.
Next is Accelerating development for massive Git repositories, from the Google Developers Blog. Seroter says Google built this to better work with Android's giant repo and open sourced it for everyone else. Android's source code is enormous: over a thousand Git repositories, one point six million files, and three hundred gigabytes for a built workspace. Traditional checkouts took more than thirty minutes, and full builds could take over an hour. The new system, called Another Build File Service, or ABFS, uses content-addressable storage across the stack to lazy-load and cache source code, workspaces, and build results. It cuts checkout times to under eight seconds and full builds to just over two minutes, and it enables parallel agentic workflows. This connects nicely with the agentic engineering levels, because fast, reproducible environments are exactly what autonomous loops need.
Then there's Put MCP Next to Your API, Not on Top of It. Seroter's take is that if your architecture supports this, it makes sense, but if you find yourself duplicating code and governance, such as access controls, then probably not. The article by Jerome Louvel argues that most teams stack an MCP server in front of an existing API server, which adds a hop, a second contract, and a second deployment. Instead, he proposes exposing REST and MCP side by side over the same flows, so the same logic serves both. He also notes that MCP is outgrowing REST and GraphQL because it has built-in discovery, elicitation, and tasks, which the older protocols don't offer standard equivalents for. Seroter's caution about governance duplication is a good checkpoint before adopting this pattern.
After that, Using ephemeral disks on Cloud Run for temporary storage. Seroter highlights this as non-AI news, where you can use these as real disks versus memory masquerading as one. By default, Cloud Run's writable container filesystem lives in memory, so every byte written counts against the container's memory limit. The generally available ephemeral disks feature gives dedicated local storage that doesn't consume container memory. The article demonstrates mounting a ten-gigabyte ephemeral disk at slash tmp on a service with only half a gigabyte of RAM. Writing a one-gigabyte file to slash tmp succeeds, while the same write to the root filesystem triggers an out-of-memory kill. This is useful for apps that need temporary file storage, and it's far cheaper than provisioning extra RAM for the same purpose.
Up next is 10 Best AI Agent Frameworks in 2026 from Builder.io. Seroter says this list has stabilized a bit, but it still feels fairly dynamic, and he wonders how many people will still build custom agents versus throw skills at managed ones. The Builder.io page also surfaces a related essay, Why are we still handing AI agents API keys?, by Alice Moore, which questions the practice of giving agents direct credentials rather than a more controlled identity layer. That question ties into the broader theme of agent governance that runs through several items in this list.
Then comes Google is a Leader in the 2026 Gartner Magic Quadrant for Cloud Operational Database Management Systems. Seroter calls it a pretty crazy showing, noting that Google as the most visionary database vendor is quite a moment. The Google Cloud blog post from Sailesh Krishnamurthy and Raj Pai explains that Google was recognized as a Leader and positioned furthest in completeness of vision. The post frames this around the agentic data cloud, highlighting Spanner and AlloyDB with native AI, vector, and graph capabilities, and pointing to top rankings in the Agentic AI Operations use case in the companion Critical Capabilities report. Seroter's reaction captures the significance for Google Cloud.
Next is AI Slop Is A-OK, which could not be retrieved. Seroter says he likes it, and his commentary is that nobody becomes an expert overnight, and we need to let people work through the journey with AI without shame. The title suggests the article argues that low-quality AI-generated output is an acceptable part of the learning process rather than something to be ashamed of. Seroter's endorsement frames it as a reminder to allow people to develop their skills incrementally.
Then there's Gemini 4 Hydrogen: a faster model for general purpose robotics. Seroter says, wow, I like it, and notes that Google Antigravity becomes a full-featured workbench for robotics work. The article from the Gemini Robotics team introduces Gemini 4 Hydrogen, a faster and more efficient model for embodied reasoning and robotic control, and launches a Robotics plugin on Google Antigravity. The plugin handles hardware bring-up, MuJoCo simulation, perception and control, and application building, so a developer can go from a robot in a box to a working application in an afternoon. Examples include a snack-sorting conveyor, a soap store that fulfills orders with two robot arms, and a chess-playing robot.
Invisible engineering teams is next. Seroter describes it as advice for teams getting lost in the shuffle because they're not on the coolest, most highly-visible projects. Anton Zaides writes that every team has a default visibility level, and DevOps, security, and platform teams often sit near the bottom. Symptoms include no credit and lots of blame, accumulating operations work, difficulty in promotion calibration sessions, and engineers feeling disconnected from the business. His advice is to connect the right pipe by targeting specific stakeholders with tailored messages, ask what would happen if you stopped doing the work, and bring stakeholders into the team's world to build relationships. Seroter's summary highlights the career risk of invisible work.
After that is Agents build things. Enterprises have to run them. Seroter says yes, primitives and runtimes still matter, but to fewer people, and he salutes those who assemble and run shared platforms. The InfoWorld piece argues that Kubernetes and infrastructure are not obsolete just because agents generate code. It quotes Linux Foundation CEO Jim Zemlin saying code is cheap but trust is expensive. Real enterprise adoption still requires secure, recoverable, affordable systems, and examples like China Merchants Bank using Kubernetes to improve accelerator utilization and lower inference costs show the primitives doing heavy work. It also covers Google's Agent Substrate and Nvidia's OpenShell as evidence that major AI players still care deeply about runtime and security boundaries.
Then there's Build full-stack Android apps with Gemini in Google AI Studio. Seroter notes there's lots of new functionality for anyone who wants to build Android apps, even without knowing all the machinery underneath. The Google Developers Blog post explains that Google AI Studio, with Gemini 4 Carbon and Hydrogen, can now generate production-ready Kotlin Android apps with full-stack persistence, sign-in, and secure Gemini API calls. It provisions Firestore and Firebase Authentication, protects API keys server-side, supports publishing directly to a personal Google Play library, and adds an HD audio-enabled emulator with one-click screenshots back to the coding agent. The goal is to let people go from prompt to app on their phone quickly.
Finally, Why Git Needs Help: Moving from Files to Systems for Source Control. Seroter calls it a compelling argument by Jason, asking why Git breaks down for agentic use and what we need instead. Jason Davenport argues that Git's file-centric, line-based diffs waste context and miss structural relationships, while agent-native source control should mutate syntax trees and code property graphs directly. He also argues that agent reasoning traces, failed hypotheses, and compiler feedback should be versioned alongside code as an execution ledger. The article cites Anthropic's Scion architecture as a real-world shift to server-centric source control that cut infrastructure overhead dramatically. It's a fitting close to a list that repeatedly circled the question of what infrastructure agents actually need.
Across the list, a clear theme emerges: agents are changing how we build, but the underlying platforms, data, and governance still matter enormously. Seroter's reading list balances enthusiasm for new agentic tools with a pragmatic nod to the people who keep shared systems running.
- Daily Reading List – November 17, 2026 (#911)
- The levels of agentic engineering (from someone who’s led the climb twice)
- Accelerating development for massive Git repositories
- Put MCP Next to Your API, Not on Top of It
- Using ephemeral disks on Cloud Run for temporary storage
- 10 Best AI Agent Frameworks in 2026
- Google is a Leader in the 2026 Gartner® Magic Quadrant™ for Cloud Operational Database Management Systems
- AI Slop Is A-OK
- Gemini 4 Hydrogen: a faster model for general purpose robotics
- Invisible engineering teams
- Agents build things. Enterprises have to run them
- Build full-stack Android apps with Gemini in Google AI Studio
- Why Git Needs Help: Moving from Files to Systems for Source Control

